JPWO2015019835A1 - Electric artificial laryngeal device - Google Patents

Abstract

Provided is an electric artificial laryngeal device capable of smoothly outputting a sound source sound adapted to a voice uttered by a user. The electric artificial laryngeal device 1 collects a uttered sound generated by adjusting the sound source sound input to the vocal tract of the user P and generates a utterance signal, and a sound collecting unit 10 Includes a signal processing unit 20 that generates a sound source signal corresponding to the utterance signal generated by and a sound source signal reproduction unit 30 that reproduces the sound source signal generated by the signal processing unit 20 and outputs a sound source sound.

Description

  The present invention outputs sound (hereinafter referred to as “sound source sound”) that serves as a sound source in its own body, such as a person who has removed the larynx including the vocal cord due to a disease such as laryngeal cancer, or a person whose vocal cord does not function normally. Electric artificial sound input from outside the body into the vocal tract (space formed by the nasal cavity, oral cavity, tongue, etc., the same shall apply hereinafter) of a person who cannot or is difficult to perform The present invention relates to a laryngeal device.

  A healthy person who is not a larynx abnormal person (hereinafter referred to as “normal larynx”) inputs sound source sound generated by vibrating the vocal cords by air discharged from the lungs and passing through the trachea into the vocal tract for articulation processing Sounds from the mouth (hereinafter referred to as “voiced sound”) by sounding (modulating by resonating the sound source sound in the vocal tract, hereinafter the same).

  However, even if the larynx abnormal person has normal articulation processing function of the vocal tract, it is impossible or difficult to generate sound source sound in the body and input to the vocal tract. Can't make utterance sound.

  Thus, an electric artificial laryngeal device that inputs sound source sound into the vocal tract of an abnormal laryngeal person by vibrating in close contact with the outside of the throat of the abnormal laryngeal person is widely used. By using this electric artificial laryngeal device, a person with abnormal larynx can input sound source sound into the vocal tract. Therefore, a person with abnormal larynx emits a desired utterance sound by a simple and easy operation of changing the shape of the vocal tract (for example, moving the mouth or tongue) in the same manner as when a normal larynx utters a sound. Is possible.

  However, the sound source sound generated by the electric artificial laryngeal device is determined irrespective of the words and utterance contents (that is, the articulation process) generated by the abnormal larynx. For example, the sound source sound emitted from the electric artificial laryngeal device may be constant without changing the fundamental frequency (pitch) with time. For this reason, it is extremely difficult for a person with abnormal larynx to add accents and intonation (for example, changes in tone due to variations in the fundamental frequency or amplitude of the sound source sound) to the uttered sound. As a result, the utterance sound produced by the person with abnormal larynx can be heard as a mechanical sound or difficult to be transmitted correctly, which is problematic.

  These problems will be specifically described with reference to FIGS. FIG. 5 is a graph showing various characteristics of vocal sounds produced by a normal larynx. FIG. 6 is a graph showing various characteristics of uttered sounds produced by an abnormal larynx using an electric artificial laryngeal device. In the graphs of FIGS. 5 and 6, signal waveforms, fundamental frequencies, non-periodic components, and spectrograms are shown as the characteristics of each uttered sound.

  5 and 6, in the signal waveform graph, the horizontal axis represents time, and the vertical axis represents amplitude. In the fundamental frequency graph, the horizontal axis represents time and the vertical axis represents frequency. In the aperiodic component graph, the horizontal axis represents time, and the vertical axis represents intensity. Further, the spectrogram indicates that the horizontal axis is time, the vertical axis is frequency, and the darker the color (closer to black), the higher the intensity.

  Of the various features of the uttered sound shown in FIGS. 5 and 6, the signal waveform indicates the overall features of the uttered sound. The fundamental frequency mainly indicates the characteristics of the sound source sound. Further, the non-periodic component mainly indicates the characteristics of the sound source sound (specifically, the timbre representing the blurred state of the uttered sound). The spectrogram shows the characteristics of articulation processing in the vocal tract.

  As shown in FIG. 5, the fundamental frequency of the sound produced by a normal larynx person changes with time and is not constant. In other words, accents and intonation are added to the utterance sound produced by a normal larynx person.

  On the other hand, as shown in FIG. 6, the fundamental frequency of the uttered sound produced by the larynx abnormal person is constant without changing over time. That is, accents and intonation are not added to the utterance sound produced by the larynx abnormal person. For this reason, the utterance sound produced by the person with abnormal larynx is heard as a mechanical sound or is difficult to be transmitted correctly.

  Therefore, Patent Document 1 proposes an electric artificial laryngeal device that controls the fundamental frequency and volume of sound source sound in accordance with myoelectric potential, joint angle, and the like detected using a sensor. Further, Patent Document 2 proposes an electric artificial laryngeal device capable of outputting a plurality of patterns of sound source sounds having different fundamental frequency variations in accordance with the operation content of a switch by an abnormal laryngeal person.

Japanese Patent Laid-Open No. 7-433 Japanese Patent Laid-Open No. 11-69476

  If the electric artificial laryngeal device proposed in Patent Documents 1 and 2 is used, it is possible to output sound sources having different fundamental frequencies. However, the electric artificial laryngeal device proposed in Patent Document 1 is output based on information (biological information obtained from a sensor attached to the outer surface of the human body) that is not directly related to the vocal sound. Since the sound source sound to be controlled is controlled, an unsuitable sound source sound may be output to the utterance sound that the larynx abnormal person wants to utter. On the other hand, the electric artificial laryngeal device proposed in Patent Document 2 needs to control the sound source sound by human operation, so that the operation of the electric artificial laryngeal device becomes complicated and the sound source sound and sound are smoothly generated. It becomes difficult to produce voice sounds.

  Therefore, an object of the present invention is to provide an electric artificial laryngeal device that can smoothly output a sound source sound suitable for a utterance sound emitted by a user.

  In order to achieve the above-mentioned object, the present invention collects a uttered sound generated by adjusting a sound source sound input to a user's vocal tract and generates a utterance signal, and the sound collecting A signal processing unit for generating a sound source signal corresponding to the utterance signal generated by the unit, and a sound source signal reproduction for reproducing the sound source signal generated by the signal processing unit and outputting the sound source sound for input to the vocal tract And an electric artificial laryngeal device characterized by comprising:

  According to this electric artificial laryngeal device, it is possible to output a sound source sound corresponding to the vocal sound actually emitted by the user.

  Further, in the electric artificial laryngeal device having the above characteristics, the signal processing unit extracts a voice feature amount indicating characteristics of the articulation processing in the user's vocal tract from the utterance signal generated by the sound collection unit. A sound source feature amount for estimating a sound source feature amount indicating a feature of a sound source sound corresponding to the articulation processing in the user's vocal tract based on the feature amount extraction unit and the voice feature amount extracted by the voice feature amount extraction unit It is preferable to include an estimation unit and a sound source signal generation unit that generates the sound source signal having the sound source feature amount estimated by the sound source feature amount estimation unit.

  According to this electric artificial laryngeal device, the sound source feature quantity estimation unit estimates the sound source feature quantity based on the voice feature quantity extracted from the utterance signal. For this reason, it is possible to accurately estimate the sound source feature amount corresponding to the articulation processing in the vocal tract while eliminating the influence of the fluctuation of the sound source sound.

  Furthermore, in the electric artificial laryngeal device having the above characteristics, the signal processing unit further includes a database in which a statistical model indicating a correspondence relationship between the audio feature quantity and the sound source feature quantity is recorded, and the sound source feature quantity It is preferable that the estimation unit estimates the sound source feature amount based on the statistical model recorded in the database.

  According to this electric artificial laryngeal device, the sound source feature amount estimation unit can easily and accurately estimate the sound source feature amount by using a statistical model built in advance.

  Furthermore, in the electric artificial laryngeal device having the above characteristics, the statistical model includes a first speech feature amount extracted from a first utterance signal generated by collecting a first utterance sound emitted by an abnormal larynx for a certain word. And the second sound source feature amount extracted from the second utterance signal generated by collecting the second utterance sound emitted by the normal larynx for the certain word, and The first voicing sound is generated after the first sound source sound input to the vocal tract of the larynx abnormal person is subjected to an articulation process, and the first voice feature amount is an articulation process in the vocal tract of the abnormal larynx person. The second voicing sound is uttered after the second sound source sound output from the vocal cords of the normal larynx is tuned in the vocal tract, and the second sound source feature amount is The characteristic of the second sound source sound Preferred.

  According to the electric artificial laryngeal device, the sound source feature amount is estimated based on the statistical model constructed using the second sound source feature amount indicating the feature of the second sound source sound output from the vocal cord of the normal larynx. . Therefore, the sound source sound output from the sound source signal reproducing unit can be brought close to a natural sound source sound output from the vocal cord of a normal larynx.

  Furthermore, in the electric artificial laryngeal device having the above characteristics, the statistical model includes a distribution of the second sound source feature amount, wherein the first sound source feature amount indicating the feature of the first sound source sound extracted from the first utterance signal is Within the range, it is preferable.

  According to this electric artificial laryngeal device, since the statistical model is constructed in a state where the first sound source feature amount and the second sound source feature amount are aligned, the sound source feature amount estimation unit calculates the sound source feature amount within the distribution range. It is possible to estimate with high accuracy.

  Furthermore, in the electric artificial laryngeal device having the above characteristics, the sound source feature amount indicates a fundamental frequency of the sound source sound, and the second sound source feature amount indicates a fundamental frequency of the second sound source sound. If there is, it is preferable.

  According to the electric artificial laryngeal device, the fundamental frequency of the sound source sound output from the sound source signal reproducing unit can be made to correspond to the articulation processing in the vocal tract.

  Furthermore, in the electric artificial laryngeal device having the above characteristics, the statistical model is based on a correspondence relationship between the first voice feature quantity and a second voice feature quantity extracted from the second utterance signal. The second voice feature is constructed by associating the first voice feature quantity with the second sound source feature quantity after correcting a shift in time direction between the first voice signal and the second voice signal. The amount preferably represents the characteristics of articulation processing in the vocal tract of the normal larynx.

  According to this electric artificial laryngeal device, even when there is a difference in speaking speed between the larynx abnormal person and the normal laryngeal person, and the time difference between the first utterance sound and the second utterance sound may occur, the deviation is corrected. The first sound feature amount and the second sound source feature amount are associated with each other. Therefore, it is possible to construct a statistical model that can accurately estimate the sound source feature amount.

  According to the electric artificial laryngeal device having the above characteristics, it is possible to output a sound source sound corresponding to the vocal sound actually emitted by the user. Therefore, it is possible to smoothly output the sound source sound that is suitable for the utterance sound emitted by the user.

The block diagram shown about the structural example of the electric artificial larynx apparatus which concerns on embodiment of this invention. The block diagram shown about the structural example of the signal processing part with which the electric artificial laryngeal apparatus shown in FIG. 1 is provided. The graph shown about an example of the construction method of a statistical model. The graph shown about an example of the construction method of a statistical model. The graph which showed about the various characteristics of the vocal sound which a larynx normal person utters. The graph which showed about the various characteristics of the utterance sound which the larynx abnormal person uses the electric artificial laryngeal device.

  First, an electric artificial laryngeal device according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of an electric artificial larynx device according to an embodiment of the present invention.

  As shown in FIG. 1, the electric artificial larynx device 1 according to the embodiment of the present invention includes a sound collection unit 10, a signal processing unit 20, and a sound source signal reproduction unit 30. In addition, in FIG. 1, the user P of the electric artificial larynx apparatus which is a larynx abnormal person other than the electric artificial laryngeal apparatus 1 is illustrated for convenience of explanation.

  The sound collection unit 10 includes, for example, an air conduction microphone, a body conduction microphone, and the like, collects a utterance sound emitted by the user P, and generates an utterance signal by converting the sound into an electric signal. At this time, the sound collection unit 10 collects uttered sound at a sampling frequency of 16 kHz, for example, and generates a uttered signal. In addition, when using a body conduction microphone as the sound collection unit 10, for example, a non-audible murmur (NAM) microphone may be used. The NAM microphone is a microphone that is used by being crimped to the back of the auricle (occipital side), and is a meat conduction microphone that collects sound propagating through the meat of the head and neck.

  The signal processing unit 20 includes an arithmetic processing device such as a CPU (Central Processing Unit) or a DSP (Digital Signal Processor), and generates a sound source signal corresponding to the utterance signal generated by the sound collection unit 10. However, the sound source signal generated by the signal processing unit 20 varies in time corresponding to the utterance signal that varies in time. For example, the sound source signal generated by the signal processing unit 20 has a fundamental frequency that can be temporally fluctuated like a sound source sound output from a normal larynx by a vocal cord (see the fundamental frequency graph in FIG. 5). .

  The sound source signal reproducing unit 30 reproduces the sound source signal generated by the signal processing unit 20 and outputs a sound source sound to be input to the user P's vocal tract. For example, the sound source signal reproducing unit 30 includes a diaphragm and a driving device for the diaphragm, and the driving device vibrates the diaphragm according to the sound source signal, and outputs sound source sound. At this time, the diaphragm is vibrated while being pressed against the throat of the user P, so that the sound source sound is input to the vocal tract of the user P.

  Then, the user P changes the shape of his or her vocal tract (for example, moves his mouth or tongue), and adjusts the sound of the sound source to produce a uttered sound. Furthermore, the utterance sound emitted by the user P is collected by the sound collection unit 10 and the series of operations described above is performed again. Thus, the electric artificial laryngeal device 1 continuously outputs the sound source sound corresponding to the uttered sound emitted by the user P and continuously inputs it to the vocal tract of the user P.

  As described above, the electric artificial laryngeal device 1 according to the embodiment of the present invention can output a sound source sound corresponding to the vocal sound actually emitted by the user P. Therefore, it is possible to smoothly output a sound source sound suitable for the utterance sound emitted by the user P.

  In the electric artificial laryngeal device 1, the sound source sound corresponding to the utterance sound that the user P is actually uttering is a certain amount of time (for example, the time required for the processing of the signal processing unit 20, etc., 50 ms to 70 ms). After that, it is input to the vocal tract of the user P. However, this time lag is negligible, and humans are insensitive to this time lag, so problems such as the listener feeling uncomfortable about the voices of the user P are unlikely to occur.

  Next, the signal processing unit 20 included in the electric artificial laryngeal device 1 shown in FIG. 1 will be specifically described with reference to the drawings. FIG. 2 is a block diagram showing a configuration example of the signal processing unit 20 included in the electric artificial laryngeal device shown in FIG.

  As shown in FIG. 2, the signal processing unit 20 includes an audio feature amount extraction unit 21, a sound source feature amount estimation unit 22, a database 23, and a sound source signal generation unit 24.

  The voice feature quantity extraction unit 21 extracts a voice feature quantity that is a feature of the articulation processing in the vocal tract of the user P from the utterance signal generated by the sound collection unit 10. The voice feature amount is based on, for example, a spectrum envelope (rough shape of frequency spectrum).

  For example, the voice feature quantity extraction unit 21 has a frame length of 25 ms and a frame shift length of 5 ms, and generates a roughly formed component from a frequency spectrum obtained by performing a short time Fourier transform (STFT) on the utterance signal. (For example, a cepstrum is obtained by performing inverse Fourier transform on a frequency spectrum in which the amplitude is logarithm, and a low-order component of the cepstrum is selectively extracted and then further Fourier transformed) Thus, the spectrum envelope is obtained continuously. Note that the spectrograms as shown in FIGS. 5 and 6 can be obtained by arranging the spectral envelopes obtained in this way continuously in the time direction. And the audio | voice feature-value extraction part 21 obtains an audio | voice feature-value by carrying out dimension compression of the spectrum envelope in the segment which couple | bonds four frames before and behind, for example with respect to each frame, for example.

  The sound source feature quantity estimation unit 22 is a sound source corresponding to the articulation processing in the vocal tract of the user P based on the voice feature quantity extracted by the voice feature quantity extraction unit 21 and the statistical model stored in the database 23. A sound source feature amount indicating a sound feature is estimated. The sound source feature amount is, for example, a fundamental frequency.

  Here, an example of a method for constructing a statistical model stored in the database 23 will be described with reference to the drawings. 3 and 4 are graphs showing an example of a statistical model construction method.

  The statistical model consists of a utterance sound (hereinafter referred to as “first utterance sound”) uttered by a person with abnormal larynx for a certain word, and a utterance sound (hereinafter referred to as “second utterance sound”) generated by a normal larynx for the certain word. Are associated with each other. The first vocal sound is generated by a person with abnormal laryngeal rhythmic processing of a sound source sound (hereinafter referred to as “first sound source sound”) output by a conventional electric artificial laryngeal device in the vocal tract. . The second utterance sound is generated by a person with normal larynx by performing a tone adjustment process in the vocal tract on a sound source sound output from the vocal cords (hereinafter referred to as “second sound source sound”).

  FIG. 3A shows an utterance signal (hereinafter referred to as “first utterance signal”) generated by collecting the first uttered sound and an utterance signal (hereinafter referred to as “first utterance signal”) generated by collecting the second uttered sound. , “Second utterance signal”). FIG. 3B is a graph showing a method of associating the first utterance signal and the second utterance signal. Note that both graphs shown in FIG. 3A and FIG. 3B are obtained when the larynx abnormal person and the larynx normal person utter the same words.

  As shown in FIG. 3 (a), even if the larynx abnormal person and the larynx normal person utter the same word, there are individual differences in the speaking speed of the person, so the first utterance signal and the second utterance signal have time. Misalignment can occur.

  Therefore, as shown in FIG. 3B, a speech feature amount extracted from the first utterance signal (hereinafter referred to as “first speech feature amount”) and a speech feature amount extracted from the second utterance signal (hereinafter referred to as “first speech feature amount”). , “Second audio feature amount”) and the time difference is corrected. As a result, a statistical model capable of accurately estimating the sound source feature amount can be constructed. The first voice feature quantity and the second voice feature quantity can be extracted by the same method as the voice feature quantity extraction method in the voice feature quantity extraction unit 21 shown in FIG.

  First, the patterns of the first voice feature quantity and the second voice feature quantity are compared, and a correspondence relationship in which the shift in the time direction is corrected using a similar feature as a clue (in FIG. 3B). (Dashed line). Then, according to the correspondence relationship, the first voice feature amount is associated with the sound source feature amount extracted from the second utterance signal (hereinafter referred to as “second sound source feature amount”). Note that the second audio feature quantity and the second sound source feature quantity are both extracted from the second utterance signal, and therefore there is no time lag between them. Further, as a method for extracting the second sound source feature amount from the second utterance signal, various known methods can be applied. For example, the method shown in Reference Document 1 below may be applied.

(Reference 1)
H. Kawahara, H. Katayose, A. de Cheveigne, and RD Patterson.
Fixed point analysis of frequency toinstantaneous frequency mapping for accurate estimation of F0 and periodicity.
Proc. EUROSPEECH, pp. 2781-2784, Budapest, Hungary, Sep. 1999.

  A statistical model is constructed by associating the first sound feature quantity and the second sound source feature quantity with respect to various words. Such a statistical model can be constructed using, for example, a mixed normal distribution model (GMM: Gaussian Mixture Model). In FIGS. 4A and 4B, for simplification of illustration and description, each of the first sound feature amount and the second sound source feature amount is a scalar. It is preferable that each of the two sound source feature amounts is a vector composed of a plurality of components because the sound source feature amount can be estimated with higher accuracy.

  The graph shown in FIG. 4A is a histogram of data of the first sound feature quantity and the second sound source feature quantity. The graph shown in FIG. 4B is a statistical model constructed by applying the GMM model to the data shown in FIG. In the graph (statistical model) shown in FIG. 4B, the higher the portion in the graph, the higher the probability of occurrence of the combination of the first sound feature amount and the second sound source feature amount.

  The sound source feature quantity estimation unit 22 estimates a sound source feature quantity based on the statistical model and the voice feature quantity extracted by the voice feature quantity extraction unit 21. At this time, if an estimation process considering the correlation in the time direction is used, the sound source feature amount estimation unit 22 can accurately estimate the sound source feature amount. Note that various known methods can be applied to the estimation processing considering the correlation in the time direction. For example, the method shown in Reference Document 2 below may be applied.

(Reference 2)
T. Toda, M. Nakagiri, K. Shikano.
Statistical voice conversion techniques for body-conducted unvoiced speech enhancement.
IEEE Transactions on Audio, Speech and Language Processing, Vol. 20, No. 9, pp. 2505-2517, Sep. 2012.

  For example, the sound source feature quantity estimation unit 22 applies the voice feature quantity extracted by the voice feature quantity extraction unit 21 to the first voice feature quantity in the statistical model to correspond (for example, the occurrence probability is maximized). Find the sound source feature. The sound source feature amount estimation unit 22 outputs the obtained second sound source feature amount as the estimated sound source feature amount.

  Finally, the sound source signal generation unit 24 has a sound source signal having the sound source feature amount estimated by the sound source feature amount estimation unit 22 (for example, if the sound source feature amount is a fundamental frequency, a sound source having a signal waveform of the fundamental frequency). Sound) is generated and output to the sound source signal reproducing unit 30 shown in FIG.

  In the signal processing unit 20, the sound source feature amount estimation unit 22 estimates the sound source feature amount based on the voice feature amount extracted from the utterance signal. For this reason, it is possible to accurately estimate the sound source feature amount corresponding to the articulation processing in the vocal tract while eliminating the influence of the fluctuation of the sound source sound.

  Further, in the signal processing unit 20, the sound source feature amount estimation unit 22 can easily and accurately estimate the sound source feature amount by using a statistical model constructed in advance. In particular, the signal processing unit 20 estimates the sound source feature amount based on a statistical model constructed using the second sound source feature amount indicating the feature of the second sound source sound output from the vocal cord of the normal larynx. Therefore, the sound source sound output from the sound source signal reproduction unit 30 can be brought close to a natural sound source sound output from the vocal cords of a normal larynx.

  Note that when the statistical model described above is constructed, the first sound source feature amount indicating the feature of the first sound source sound extracted from the first utterance signal may be within the distribution range of the second sound source feature amount. Good. In this way, the statistical model is constructed in a state where the first sound source feature value and the second sound source feature value are aligned (for example, based on the first utterance signal and the second utterance signal that are both masculine, Since a statistical model is constructed, or a statistical model for women is constructed based on both the female first utterance signal and the second utterance signal), the sound source feature quantity estimation unit 22 is within the distribution range. This is preferable because it is possible to accurately estimate the sound source feature amount.

  For example, in this case, first, a sound source feature amount desired by the user P (hereinafter referred to as “target sound source feature amount”) is determined. Specifically, for example, the pitch (basic frequency) desired by the user P is determined. Then, it is possible to extract the first utterance signal that can extract the first sound source feature amount that matches or approximates the target sound source feature amount, and the second sound source feature amount that matches or approximates the target sound source feature amount. After obtaining each of the second utterance signals, a statistical model is constructed according to the method described above.

  The first utterance signal as described above collects the first utterance sound produced by the user P or the like using the electric artificial laryngeal device whose output is adjusted so that the first sound source feature amount is obtained. Can be obtained. The first utterance signal as described above is also obtained by adjusting the first utterance signal already recorded in the database or the like so that the first utterance feature amount approaches the target utterance feature amount. be able to. A statistical model with a wide distribution range of the first sound source feature amount by simultaneously using various first utterance signals obtained by adjusting the output of the electric artificial laryngeal device and adjusting the first sound source feature amount. May be built.

  In addition, the second utterance signal as described above selects a normal larynx person having a vocal cord from which the second sound source feature amount can be obtained, and collects a second utterance sound emitted by the normal larynx person. Can be obtained. The second utterance signal as described above is also acquired by adjusting the second utterance signal already recorded in the database or the like so that the second utterance feature amount approaches the target utterance feature amount. be able to.

<Deformation, etc.>
[1] In the embodiment of the present invention described above, the electric artificial laryngeal device 1 mainly outputs the sound source sound by varying the fundamental frequency so as to correspond to the vocal sound (particularly, articulation processing in the vocal tract). As explained. However, the amplitude (power) of the sound source sound may be varied so as to correspond to the uttered sound, or both the fundamental frequency and amplitude of the sound source sound may be varied and output so as to correspond to the uttered sound. Also good.

  When the electric artificial laryngeal device 1 is configured to output not only the fundamental frequency of the sound source sound but also the amplitude, accents and intonation are often added due to the fluctuation of the fundamental frequency of the sound source sound. It is not limited to languages (for example, Japanese), and it is possible to output sound sources for various languages, such as languages that often add accents and intonation due to variations in the amplitude of the sound source (for example, English) It becomes.

[2] The electric artificial laryngeal device 1 is configured to switch the presence or absence of the output of the sound source sound according to the behavior of the user P (for example, presence or absence of an operation such as pressing a button or pressing the main body against the throat). Is preferred.

  In this case, a sound source sound having a predetermined sound source characteristic amount is output for a very short time at which the electric artificial larynx device 1 starts outputting the sound source sound. However, since a sound source sound corresponding to the utterance sound of the user P is output immediately thereafter, problems such as a listener feeling uncomfortable about the utterance sound of the user P hardly occur.

[3] For example, a person whose vocal cords will not function in the future (ie, a person who will become the above-mentioned user P in the future), such as a patient who is scheduled to have the larynx removed by surgery, During that time, it is preferable to collect and record a vocal sound using the person's own vocal cord (hereinafter referred to as a “personal vocal cord vocal sound”).

  The recorded personal vocal cord utterance is a utterance produced by a normal larynx, and is included in the second utterance. Therefore, it is preferable to construct a statistical model using the second vocal sound including the vocal cord vocal sound. Further, since the sound source feature amount extracted from the signal of the vocal cord vocal sound is considered to be the sound source feature amount desired by the user P, when the statistical model is constructed using the sound source feature amount as the above-described target sound source feature amount, ,preferable.

  The statistical model constructed in this way reflects the utterance characteristics (accent, intonation, etc.) when the user P is a normal larynx. Therefore, by using this statistical model in the electric artificial laryngeal device 1 described above, it is possible to output a sound source sound that effectively reproduces the characteristics of the utterance when the user P is a normal larynx. Become.

  It should be noted that the greater the amount of recording of the vocal cord vocal sound, the better, but it may be about 50 sentences (the amount that requires about 3 to 5 minutes to read out).

  INDUSTRIAL APPLICABILITY The present invention can be suitably used for an electric artificial laryngeal device that inputs sound source sound into the vocal tract of an abnormal larynx.

DESCRIPTION OF SYMBOLS 1: Electric artificial larynx apparatus 10: Sound collection part 20: Signal processing part 21: Voice feature-value extraction part 22: Sound source feature-value estimation part 23: Database 24: Sound source signal generation part 30: Sound source signal reproduction part P: User

Claims (7)

A sound collection unit that collects the utterance sound generated by the articulation processing of the sound source sound input to the user's vocal tract, and generates an utterance signal;
A signal processing unit that generates a sound source signal corresponding to the utterance signal generated by the sound collection unit;
A sound source signal reproducing unit that reproduces the sound source signal generated by the signal processing unit and outputs a sound source sound for input to the vocal tract;
An electric artificial laryngeal device comprising: The signal processing unit is
A voice feature amount extraction unit that extracts a voice feature amount indicating characteristics of articulation processing in the user's vocal tract from the utterance signal generated by the sound collection unit;
A sound source feature amount estimation unit that estimates a sound source feature amount indicating a feature of a sound source sound corresponding to articulation processing in the user's vocal tract based on the sound feature amount extracted by the sound feature amount extraction unit;
A sound source signal generation unit that generates the sound source signal having the sound source feature amount estimated by the sound source feature amount estimation unit;
The electric artificial laryngeal device according to claim 1, comprising: The signal processing unit further comprises a database in which a statistical model indicating a correspondence relationship between the audio feature quantity and the sound source feature quantity is recorded;
3. The electric artificial laryngeal device according to claim 2, wherein the sound source feature amount estimation unit estimates the sound source feature amount based on the statistical model recorded in the database. The statistical model includes a first speech feature amount extracted from a first utterance signal generated by collecting a first utterance sound produced by a person with abnormal larynx for a certain word, and a first utterance produced by a normal larynx person for the certain word. It is constructed by associating the second sound source feature amount extracted from the second utterance signal generated by collecting two utterance sounds,
The first uttered sound is generated after the first sound source sound input to the vocal tract of the larynx abnormal person is subjected to articulation processing,
The first audio feature amount indicates a characteristic of articulation processing in the vocal tract of the larynx abnormal person,
The second vocal sound is generated by the second sound source sound output from the vocal cord of the normal larynx being subjected to articulation processing in the vocal tract,
The electric artificial laryngeal device according to claim 3, wherein the second sound source feature amount indicates a feature of the second sound source sound.   5. The statistical model is characterized in that a first sound source feature amount indicating a feature of the first sound source sound extracted from the first utterance signal is within a distribution range of the second sound source feature amount. The electric artificial laryngeal device described in 1.   6. The sound source feature amount indicates a fundamental frequency of the sound source sound, and the second sound source feature amount indicates a fundamental frequency of the second sound source sound. The electric artificial laryngeal device as described. The statistical model is based on a correspondence relationship between the first voice feature quantity and the second voice feature quantity extracted from the second voice signal, and the time direction of the first voice signal and the second voice signal Is constructed by associating the first sound feature quantity with the second sound source feature quantity after correcting the shift in
The electric artificial laryngeal device according to any one of claims 4 to 6, wherein the second voice feature amount indicates a feature of articulation processing in the vocal tract of the normal larynx.

Images